Beyond Linear Horizons: NN and RL Algorithms Eclipse Traditional Models in High-Fidelity Simulation and Control for Omnidirectional Robot

Omnidirectional swerve drivetrains in the FIRST Robotics Competition (FRC) demand high-fidelity simulations and optimized motion profiles for peak performance. This study compares linear models-DC motor feedforward and state-space representations-against nonlinear NN architectures, including FNN, LSTM and TCN models for motor motion prediction. It also evaluates traditional trapezoidal motion profiles versus RL-based profiles using PPO. We prototyped a functional FRC swerve module, calibrated and deployed it to collect high-resolution motion data. The cleaned voltage and velocity data from each driving motor were accurately and efficiently predicted by traditional linear models, yet the mechanics of steering motors, which were more prone to friction and nonlinear forces, were not well captured by linear models. Nonetheless, our experimented nonlinear models showed great improvement from the linear benchmark. In motion profiling, a PPO-RL framework with curriculum learning for a straight path yielded policies that reduced average completion time of the robot compared to that of a traditional trapezoidal profile. This data-driven data collection, model training, and RL deployment might offer FRC teams accessible tools for optimization, with extensions to broader wheeled robotics applications. GitHub Repository: https://github.com/yurandai/FRC_Swerve_ML_Experiments